Software and AI patents in Europe

What Article 52 EPC actually excludes

Article 52(2) EPC lists things that are not "inventions": discoveries, scientific theories and mathematical methods; aesthetic creations; schemes, rules and methods for performing mental acts, playing games or doing business, and programs for computers; and presentations of information. Article 52(3) then limits the exclusion to these things "as such" (our annotated Article 52 page collects the key Board decisions).

Those two words carry the whole system. A program is excluded only when it lacks technical character. Since T 1173/97 (IBM, 1998), a program has technical character if, when run, it produces a "further technical effect" going beyond the normal physical interaction between software and hardware, such as electrical currents in the processor (Guidelines G-II 3.6). Controlling an anti-lock braking system, compressing video, restoring a distorted image and encrypting communications all qualify, as do programs designed around the internal functioning of the computer, such as memory allocation or protection against power-analysis attacks (G-II 3.6.1).

For claims to a computer-implemented method, a programmed device or a computer-readable medium, the question does not even arise. Any method using technical means, and any technical means itself, has technical character (T 258/03, T 424/03, G 3/08). The EPO calls these "computer-implemented inventions": claims where at least one feature is realised by a computer program.

The two hurdles: any hardware, then COMVIK

EPO examination of software runs in two stages, and almost every refusal happens at the second.

1. Is it an invention at all?

In T 258/03 (Hitachi, 21 April 2004), a Board held that a method involving technical means is an invention under Article 52(1). A computer is enough. This is the "any hardware" test. It is a low bar by design, and the Enlarged Board endorsed it in G 1/19.

2. Is there an inventive technical contribution?

T 641/00 (COMVIK, 26 September 2002) set the rule for mixed inventions: inventive step is assessed taking account of all features that contribute to the technical character, while features making no such contribution cannot support an inventive step. The Guidelines turn this into a version of the problem-solution approach (G-VII 5.4):

1. Identify which features contribute to technical character, based on the technical effects they achieve in the context of the invention.
2. Pick the closest prior art, focusing on those technical features.
3. Find the differences. If none of them makes a technical contribution, the claim lacks inventive step even if it is new.
4. Formulate the objective technical problem from the technical effects. Non-technical features can be handed to the skilled person as a constraint, like a requirements specification from the business side (G-VII 5.4.1).
5. Ask whether the technical solution would have been obvious to the skilled person, usually a software engineer or a specialist in the application field.

The practical consequence: if your novelty lies in a pricing rule, a recommendation policy or a new game mechanic, the examiner hands that rule to a programmer as a given and asks only whether implementing it was obvious. It usually is. The Guidelines' own example is an online game that reduces network traffic by capping the number of players. The traffic saving comes from changing the rules, which circumvents the technical problem rather than solving it, so it cannot ground an inventive step (T 258/03, T 414/12).

What counts as a technical effect

A mathematical method, and by extension an algorithm, contributes to technical character in two ways (G-II 3.3, citing T 2330/13):

• Technical application: it serves a specific technical purpose, and the claim is functionally limited to that purpose, with the inputs and outputs causally linked to the technical effect.
• Technical implementation: it is specifically adapted to the hardware it runs on, its design motivated by technical considerations about the internal functioning of the computer or network.

Two points catch people out. First, naming the type of data is not enough: a claim that processes "sensor data" or "image data" without saying what the result is used for may still lack a technical purpose (T 2035/11, T 1029/06, T 1161/04). Second, being faster is not a technical effect on its own. An algorithm that is more efficient than earlier ones but runs on a generic computer for a non-technical purpose contributes nothing (G-II 3.3, T 1370/11). Once a technical effect is established, though, computational efficiency in achieving it does count.

AI and machine learning in the 2026 Guidelines

The EPO's position on AI sits in G-II 3.3.1 of the April 2026 Guidelines, as a subsection of mathematical methods. That placement tells you how the examiner starts. Neural networks, genetic algorithms, support vector machines, k-means, kernel regression and discriminant analysis are "per se of an abstract mathematical nature, irrespective of whether they can be trained". The general rules of G-II 3.3 then apply, and the main points are these:

• Words are not hardware. "Neural network", "support vector machine" or "reasoning engine" may refer only to an abstract model and do not by themselves imply technical means. Claim the computer, the sensor or the processor explicitly.
• Purpose decides. A neural network in a heart monitor that identifies irregular heartbeats makes a technical contribution, and so does classifying images, video, audio or speech from low-level features such as edges or pixel attributes. Classifying text by its content is a linguistic purpose, not a technical one (T 1358/09). Classifying abstract data records with no technical use of the result is not technical either, however good the algorithm's mathematical properties (T 1784/06).
• Training can count. Where the classifier serves a technical purpose, the steps of generating the training set and training the model can also contribute to technical character if they help achieve that purpose.
• Hardware adaptation counts. The Guidelines' example is assigning data-intensive training steps to a GPU and preparatory steps to a CPU to exploit the parallel architecture (G-II 3.3). The claim has to recite the GPU and CPU split.
• Show the effect. The effect can be established by explanation, mathematical proof or experimental data. Mere assertion is not enough, but comprehensive proof is not required.
• Training data disclosure. If the effect depends on particular characteristics of the training data, the application must disclose the characteristics needed to reproduce it, unless the skilled person can determine them without undue burden. The dataset itself generally need not be disclosed.

G 1/19 and simulations

The Enlarged Board's decision in G 1/19 (Pedestrian simulation, 10 March 2021) is the leading modern authority. It held that a computer-implemented simulation of a technical system, claimed as such, can solve a technical problem by producing a technical effect going beyond its implementation on a computer. But it is not enough that the simulation is based on technical principles of the simulated system, and the answer is the same when the simulation is claimed as part of a design process.

G 1/19 matters well beyond simulations. It confirmed the any hardware approach and the COMVIK framework, and the Guidelines cite it for the rule that technical effects must be achieved over the whole scope of the claim: substantially all embodiments must show the effect (G-VII 5.4.1). A broad claim to "a machine learning model" that only sometimes produces the effect is vulnerable for that reason. For simulations, output data with only a "potential technical effect" is considered only if the intended technical use is explicitly or implicitly specified in the claim, for example because the output is control data for a technical device (G-II 3.3.2). Inputs or outputs linked to physical reality, such as indirect measurement of a real object's state, are on firmer ground.

Worked example: drafting an ML claim that survives

Suppose the invention is a convolutional neural network that detects early bearing faults in wind-turbine gearboxes. An accelerometer on the bearing housing produces vibration signals; these are converted to spectrograms; the network, trained on spectrograms labelled with known bearing conditions, outputs a fault class; and the controller reduces rotor speed when a fault is detected. Here is the same invention claimed three ways.

Draft A: refused

1. A computer-implemented method comprising: receiving input data; processing the input data with a trained neural network to obtain a classification; and outputting the classification.

It passes the any hardware test, because it is computer-implemented. It fails at inventive step. The data are abstract, the purpose is unstated, and the neural network is a mathematical model. Under COMVIK the only technical feature left is a generic computer, and running a classifier on a computer was obvious long ago.

Draft B: still at risk

1. A computer-implemented method comprising: receiving vibration data; classifying the vibration data with a trained convolutional neural network; and outputting a fault class.

Better, but naming the data type does not necessarily make the method technical (G-II 3.3). Vibration of what? Measured how? Used for what? The examiner can argue that the claim covers classifying any dataset labelled "vibration", that the effect is not achieved across the whole scope, and that the link between the maths and a technical purpose is missing.

Draft C: a claim with a real technical contribution

1. A computer-implemented method of detecting a fault in a rolling-element bearing of a wind-turbine gearbox, comprising: receiving a vibration signal measured by an accelerometer mounted on a housing of the bearing; converting the vibration signal into a time-frequency spectrogram; inputting the spectrogram into a convolutional neural network trained on spectrograms labelled with bearing conditions, to determine a condition class of the bearing; and, when the condition class indicates a fault, generating a control signal that reduces the rotational speed of the turbine rotor.

Now every step is tied to a technical purpose. Deriving the physical state of a real object from measurements is an indirect measurement, which the Guidelines treat as a technical contribution regardless of what is done with the result (G-II 3.3). The control step adds a second technical effect. The network and its training now contribute to technical character because they serve that purpose (G-II 3.3.1), so the examiner must weigh them in the inventive step analysis instead of discarding them.

Technical character is not inventive step. Draft C still has to be non-obvious over the closest prior art: if earlier turbines already ran a CNN on bearing spectrograms, you need a further technical difference, such as the spectrogram resolution chosen for the bearing's fault frequencies, with test data in the description showing earlier detection. The description should also state what the training data must contain (sensor placement, sampling rate, how conditions were labelled) so that the effect can be reproduced.

To complete the set, add the claim forms the Guidelines accept without a Rule 43(2) objection: a data processing apparatus comprising means for carrying out the method, a computer program comprising instructions which, when executed by a computer, cause it to carry out the method, and a computer-readable storage medium carrying those instructions (F-IV 3.9.1). Our guide on how to write patent claims covers claim structure in more depth.

Drafting checklist for software and AI applications

• Write the technical problem into the description. Since G 1/24 (18 June 2025), the description and drawings are always consulted to interpret the claims. Explain the effect, how the claimed features cause it, and give data where you have it.
• Keep a fallback position. Put the hardware-specific features (sensor type, GPU allocation, memory layout, network protocol) in dependent claims and in the description. They are what you amend to when the examiner sets aside the business logic.
• Avoid result-only features. "Wherein the accuracy is improved" is a desired result, not a technical feature. Say how.
• Separate business and technical effects. "Increases sales" and "better recommendations" are not technical effects. "Reduces the bandwidth needed to transmit the model update" might be.
• Check the art before you file. Under COMVIK, the closest prior art is chosen with the technical features in mind, so search for the technical implementation, not just the business idea.

The filing route, fees and timeline are the same as for any other invention; see the European patent application guide. Once granted, a software patent can be turned into a Unitary Patent like any other European patent.

The UK position after Emotional Perception

For almost 20 years the UK applied its own test from Aerotel Ltd v Telco Holdings Ltd [2006] EWCA Civ 1371, a four-step approach asking whether the claim made a novel technical contribution, ignoring contributions that fell within an exclusion. The EPO's Boards criticised it soon after, in T 154/04 (Duns), for mixing up the "invention" question with novelty and inventive step.

On 11 February 2026 the UK Supreme Court settled the question in Emotional Perception AI Ltd v Comptroller General of Patents, Designs and Trade Marks [2026] UKSC 3. The application concerned an artificial neural network that recommends media files likely to create a similar emotional response. According to the court's press summary, it unanimously allowed the applicant's appeal and held that:

• the Aerotel approach should no longer be followed, and UK courts should follow the approach endorsed by the Enlarged Board in G 1/19, including the any hardware test;
• an artificial neural network is a "program for a computer", interpreted as a set of instructions that any kind of computer can follow to manipulate data;
• but the claims were not to a program "as such", because the network can only be implemented on hardware, so the UKIPO was wrong to refuse the application on that ground; and
• the next stage is G 1/19's "intermediate step", filtering out features that do not contribute to technical character before novelty and inventive step. The court declined to apply it itself and sent the case back to the Hearing Officer.

The court also said that UK courts are not obliged to use the EPO's problem-solution method for inventive step. The UKIPO published a practice notice on 14 July 2026 explaining how examiners now apply section 1 of the Patents Act 1977. In short, a software claim that would be refused at the EPO for lack of technical contribution should not now expect a different result in the UK; the UK simply reaches it by a route closer to the EPO's.

Europe and the US compared

In the US the question is subject matter eligibility under 35 U.S.C. 101. In Alice Corp. v. CLS Bank International, 573 U.S. 208 (decided 19 June 2014), the Supreme Court applied a two-step test: is the claim directed to an abstract idea, and if so, does it add an "inventive concept" that transforms it? Claims that merely require generic computer implementation of an abstract idea fail. The USPTO applies Alice through the framework in MPEP 2106, where a claim that integrates the abstract idea into a practical application survives at Step 2A, prong 2.

For AI, the most important recent USPTO authority is Ex parte Desjardins, Appeal 2024-000567, an Appeals Review Panel decision of 26 September 2025 that the USPTO designated precedential on 4 November 2025. The USPTO lists it as concerning a technological improvement to a machine learning model under Step 2A, prong 2 (USPTO precedential decisions). That is a real difference from the EPO, where an improvement to a model with no technical application and no hardware-specific implementation is still treated as mathematics. If you file in both places, draft the description to support both arguments, and see our guide to answering a section 101 rejection for how the US argument is made in practice.

More EPO guides

• How to file a European patent application

  Direct or Euro-PCT filing, search, examination, grant and validation, with the EPO fees in force from 1 April 2026.

• Patent drawing requirements: USPTO, EPO and India

  37 CFR 1.84, the EPO rules that replaced Rule 46 EPC, and Indian rule 15, compared side by side.

• How to write patent claims: USPTO, EPO and India

  Claim structure and worked examples, then the claim rules and fees at the USPTO, the EPO and the Indian Patent Office, side by side.

• The Unitary Patent and the UPC, explained

  The 18 countries covered, renewal fees, the UPC opt-out, and when classic validation is the better choice.

• EPO fees 2026: the full fee schedule

  Every EPO fee in force from 1 April 2026, the 30% micro-entity reduction, PCT fees at the EPO and a fee calculator.

• How much does a European patent cost?

  EPO fees to grant, translations, validation and renewals, with the EPO's Unitary Patent cost comparison and two worked examples.

• European patent validation: countries and deadlines

  Translation rules and deadlines in all 40 EPC states, the London Agreement, extension and validation states, and the Unitary Patent.

• The problem-solution approach to inventive step at the EPO

  Closest prior art, the objective technical problem and the could-would test, with a fully worked example and the usual mistakes.

• European patent opposition at the EPO

  The 9-month deadline and fee, the grounds, procedure and oral proceedings, appeal to the Boards, and how it compares with UPC revocation.

• Supplementary protection certificates (SPCs) in Europe

  Who can get an SPC, how long it lasts, the filing deadline, the paediatric extension, the manufacturing waiver and the unitary SPC proposal.

• The European search report and EESR, explained

  X, Y and A citation categories, the search opinion, Rule 62a and 63 limited searches, and the Rule 70a deadline to reply.

• Entering the European phase from a PCT application

  The 31-month deadline, Rule 159 acts and fees, the supplementary search, Rule 161, translations and a worked example.

• European divisional applications under Rule 36 EPC

  When the parent is pending, fees by generation, back renewal fees, the Article 76 added-matter trap and timing.

• How to speed up EPO prosecution: PACE, PPH and waivers

  PACE since February 2026, the Rule 70(2) and Rule 161 waivers, PPH with the USPTO, and fast Rule 71(3) approval.

• The European Patent Convention, annotated

  Key EPC Articles and Rules explained one by one, from novelty and claims to Rule 71(3), with deadlines, fees and Enlarged Board case law.

• Types of patent searches, and which one you need

  Patentability, freedom-to-operate, invalidity and landscape searches compared by purpose, timing, scope and deliverable.

• Patentability search: how to check novelty before you file

  What counts as prior art under 35 U.S.C. 102 and Art. 54 EPC, the grace period gap, search steps and how to read results.

• Freedom-to-operate search in the US and Europe

  In-force claims only, legal status checks at the USPTO, EPO and UPC, FTO opinions after Halo, and design-arounds.

• Invalidity search: prior art to challenge a patent

  IPR, PGR, EPO opposition and UPC revocation deadlines, prior art dates, standards documents and claim charts.

• Patent landscape analysis, step by step

  Building a CPC and keyword dataset, cleaning families and names, the metrics that matter and the 18-month publication lag.

• How to search Espacenet

  Smart and Advanced search, field identifiers, proximity and truncation, CPC, INPADOC families, legal events and Global Dossier.

• Google Patents search guide

  Search operators and fields, Boolean and proximity syntax, non-patent literature, and where Google Patents falls short of the official databases.

• The Cooperative Patent Classification (CPC), explained

  How CPC symbols are built, CPC against IPC, the Y section, how to find the right class and how to search it in each database.

• Patent search Boolean operators and search strings

  Boolean, proximity and truncation syntax compared across Patent Public Search, Espacenet, Google Patents and The Lens, with semantic search for contrast.

• The PCT application: timeline, fees and national phase

  The international phase month by month, receiving Office and ISA fees in CHF, USD and EUR, national phase deadlines by office, and a deadline calculator.

• How much does a patent cost?

  USPTO fees by entity size from filing to the last maintenance fee, what drives professional fees, EPO fees, and an estimator.

• How long does a patent last?

  US utility and design terms, patent term adjustment and extension, terminal disclaimers, maintenance fee deadlines, Europe and India, with an expiry estimator.

• Can AI write a patent application?

  What AI drafting does well and badly, the USPTO and EPO positions on AI tools, AI inventorship after 2025, and confidentiality.

All guides, by patent office